Evaluation of Growth Differentiation Factor 15 as an Early Biomarker of Anthracycline‐Induced Cardiotoxicity Using an Integrated PK/PD Model Authors: Mohamed S. Dabour, Anne H. Blaes, Pamala A. Jacobson, Mark N. Kirstein, Beshay N. Zordoky, Shen Cheng Published: 2026 ABSTRACT ABSTRACTDoxorubicin (DOX) is an effective chemotherapeutic agent; however, its use is limited by dose‐dependent cardiotoxicity, which is challenging to detect early with current biomarkers such as NT‐proBNP. Growth differentiation factor‐15 (GDF15) has emerged as a promising cytokine associated with cardiovascular disease. This study aimed to develop a pharmacokinetic/pharmacodynamic (PK/PD) model describing the dynamic relationship between DOX and GDF15 in patients with breast cancer, compare GDF15 responses with NT‐proBNP, and evaluate the prognostic utility of GDF15. Using plasma samples from 17 breast cancer patients who received two consecutive DOX cycles, the DOX‐GDF15 PK/PD model was estimated in Monolix, using DOX empirical Bayes estimates from our published DOX‐NT‐proBNP PK/PD model. Simulations were then performed to compare exposure metrics (C max, AUC) for GDF15 and NT‐proBNP and to evaluate cardiovascular risk stratification based on predefined GDF15 thresholds. Our DOX‐GDF15 model adequately described the observed GDF15 profiles, with an indirect stimulation response model incorporating four transit compartments. Baseline GDF15 was positively associated with cumulative DOX dose and BMI. Simulations showed that GDF15 rose rapidly, peaking at ~8 h postinfusion, whereas NT‐proBNP peaked at 34 h. Correlations between GDF15 and NT‐proBNP C max and AUC were consistently positive but weak to moderate. Simulations demonstrated a cumulative DOX dose‐dependent increase in GDF15 levels and enabled cardiovascular risk stratification based on predefined GDF15 thresholds, with 8.7% of simulated patients reaching high risk, consistent with reported cardiotoxicity incidence. This PK/PD model generates the hypothesis that GDF15 may serve as an early biomarker of DOX‐induced cardiotoxicity and provides a preliminary framework for risk stratification, though further validation in larger trials is warranted. Introduction Anthracyclines such as doxorubicin (DOX) are highly effective chemotherapeutic agents widely used in breast cancer and several other malignancies [1]. However, the clinical utility of DOX is significantly limited by its dose‐dependent cardiotoxicity, which can lead to severe and irreversible cardiac dysfunction [2]. This adverse effect, often referred to as doxorubicin‐induced cardiotoxicity (DIC) or anthracycline‐induced cardiotoxicity, can manifest either acutely, or months to years after stopping treatment [2]. The risk of cardiomyopathy increases with cumulative anthracycline doses. Patients receiving high cumulative doses of DOX (e.g., ≥ 250–300 mg/m2) are at significantly high risk, although even lower doses can cause subclinical cardiac dysfunction [2]. Overall, an estimated 2%–9% of patients treated with anthracyclines develops cardiotoxicity [3]. Early detection of cardiac injury during anthracycline therapy is critical to enable timely interventions before the onset of overt and irreversible heart failure. Current strategies for anthracycline‐induced cardiotoxicity monitoring rely on echocardiography assessment of left ventricular ejection fraction (LVEF) and a few circulating biomarkers of cardiac injury, among which troponin and natriuretic peptides are the most established [3, 4]. However, echocardiography may detect cardiac dysfunction only after substantial myocardial damage has occurred [5]. Similarly, the concentration of troponin, a protein released during cardiomyocyte injury, often rises during anthracycline therapy, but persistent posttreatment troponin elevations are associated with subsequent declines in LVEF [6]. The natriuretic peptide, N‐terminal pro–B‐type natriuretic peptide (NT‐proBNP) is a hormone released in response to ventricular wall stress, and is widely used for the diagnosis and risk stratification of heart failure [6]. However, NT‐proBNP has several limitations with its predictive utility in anthracycline‐induced cardiotoxicity has been mixed. Some studies showing inconsistent or transient changes during treatment, and early elevations do not always progress to clinically evident cardiomyopathy [3]. These limitations drive interest in additional biomarkers that better signal subclinical cardiac injury. One of the emerging biomarkers is growth differentiation factor‐15 (GDF15), a stress‐responsive cytokine in the transforming growth factor‐β (TGF‐β) superfamily [7]. GDF15 is upregulated in response to inflammation, oxidative stress, and tissue injury [7]. Elevated circulating levels of GDF15 have been strongly linked to adverse outcomes in several cardiovascular diseases, including atrial fibrillation [8], coronary artery disease [9] and heart failure [10, 11], supporting its value as a prognostic marker. Recently, GDF15 has gained attention as a potential early biomarker of cancer therapy‐related cardiotoxicity induced by anti‐HER2 agents [12] and DOX [4]. The present study evaluated GDF15 as a prognostic biomarker for DOX‐induced cardiotoxicity, by (1) developing a pharmacokinetic/pharmacodynamic (PK/PD) model that characterizes the dynamic interactions between DOX exposure and GDF15 concentrations in patients with breast cancer, and integrating this model with our previously developed PK/PD model involving NT‐proBNP [13]; (2) comparing the model‐simulated individual post‐DOX exposures of GDF15 against NT‐proBNP in the same patient population; and (3) assessing the prognostic performance of GDF15 as a biomarker for DOX‐induced cardiotoxicity using population simulations. Methods Study Population, Treatment, and Design This study is a secondary analysis using samples collected from 17 patients with breast cancer in a randomized, crossover clinical trial. The parent study prospectively investigated the drug–drug interaction between DOX and enalapril, an ACE inhibitor, used to mitigate anthracycline‐induced cardiotoxicity and details regarding study design, population, treatment, and sample collection were previously published [14] and summarized in the Supporting Information. PK/PD Model Development The PK model for DOX and its active metabolite, doxorubicinol (DOXOL), was extracted from our previous study [13]. Briefly DOX and DOXOL PK is characterized using a three‐compartment and a one‐compartment PK model with first‐order eliminations, respectively. The DOXOL metabolite formation was described by a first‐order metabolite formation from DOX central compartment (Figure 1A). We translated our previously published DOX/DOXOL PK model into Monolix, and individual PK parameters (i.e., empirical Bayes estimates, EBEs) were derived without changes in PK model structure and population parameters. Similarly, the previously developed NT‐proBNP PD model was translated into Monolix to obtain EBEs without modifications to the PD model structure or population parameters. Schematic representation of the pharmacokinetic–pharmacodynamic (PK/PD) models describing doxorubicin, its metabolite doxorubicinol, Nt‐proBNP, and GDF15. Figures (A) and (B) summarize the previously published PK/PD model of dox, doxol, and Nt‐proBNP [13]. (A) The pharmacokinetic component includes a three‐compartment model for doxorubicin (central compartment, V c, and two peripheral compartments, Vp2 and Vp3) with metabolism to doxorubicinol (V doxol). The pharmacodynamic component links predicted doxorubicin concentrations with indirect response models to (B) NT‐proBNP (5‐transit compartments) and (C) GDF15 (4‐transit compartments). [+]: Stimulatory effect; DOX, doxorubicin; doxol, doxorubicinol; k out, degradation rate constant; k tr, transit rate constant; Transit, transit compartment; GDF15, growth differentiation factor 15; R 0, baseline concentration; NT‐proBNP: N‐terminal probrain natriuretic peptide. Created using BioRender.com. A sequential modeling strategy was adapted for the development of GDF15 PK/PD model, where EBEs were incorporated as regressors (i.e., fixed individual parameters) to drive the GDF15 PK/PD model development [15]. Both DOX and DOXOL concentrations were tested as input to drive GDF15 stimulations in the PK/PD model, given the known cardiotoxic potential of DOXOL [16]. The GDF15 PD profiles were characterized using an indirect response model with a stimulatory effect on GDF15 production induced by either DOX or the summation of DOX and DOXOL concentrations. To account for the temporal delay between DOX exposure and the observed increase in circulating GDF15, a transit compartment structure was incorporated as empirical delay elements. A sensitivity analysis evaluating different numbers of transit compartments was conducted, and four transit compartments were found to better characterize the observed biological delays in GDF15 elevations following DOX administration. The GDF15 PD model structure is illustrated in Figure 1C. The stimulation of the production rate by DOX or DOXOL was modeled as an Emax model. The structural model is shown in Equation (1): (1) dTrans1dt=Kin×1+Emax.CdrugEC50+Cdrug−Ktr×Trans1 dTrans2dt=Ktr×Trans1−Trans2 dTrans3dt=Ktr×Trans2−Trans3 dTrans4dt=Ktr×Trans3−Trans4 dGDF15dt=Ktr×Trans4−Kout×GDF15 where K tr denotes the first‐order transit rate constant, and K out represents the degradation rate constant of GDF15, and K in is the zero‐order production rate constant, E max is the maximal stimulatory effect, EC50 is the drug concentration producing 50% of E max, C drug is the predicted plasma concentration of DOX (C DOX) or the sum of the predicted concentrations of DOX and DOXOL (C DOX + C DOXOL). Trans1–Trans4 are four sequential transit compartments used to describe the delayed response between DOX exposure and the observed increase in GDF15 concentrations. The level of GDF15 and transit compartments at baseline without perturbation were assumed to be at steady state. The levels of which are calculated using Equation (2). (2) Trans1t0=Trans2t0=Trans3t0=Trans4t0=KinKtr GDF15t0=R0=KinKout where R 0 is the baseline concentration of GDF15. Random‐Effect Model The inter‐individual variability (IIV) and inter‐occasion variability (IOV) on PK/PD parameters were modeled assuming a log‐normal distribution as follows: (3) Pi=Ppop.exp(ηi) ηi~N0ω2 where P i is the parameter estimate for the ith individual. P pop is the typical value of the parameter P. η i is a random effect for the ith individual, accounting for the interindividual differences between P i and P pop, the distribution of which was assumed to follow a normal distribution with a mean of zero and variance ω 2. The addition of IIV and IOV on a model parameter was guided by changes in objective function value (OFV) and improvement in goodness‐of‐fit diagnostics. Several residual error models (proportional, additive, and combined) were tested to characterize residual unexplained variability (RUV) [17]. The extent of η‐shrinkage was calculated for each random effect to assess the EBEs' reliability [18]. Covariate Model Potential covariates (patient‐specific characteristics listed in Table 1) were initially selected based on their clinical relevance to affect GDF15 responses following DOX administration. In addition to these covariates, cumulative DOX dose and concomitant enalapril use were also evaluated as potential covariates in GDF15 PD model. Because missing values were minimal (< 6%), missing covariate values (e.g., baseline LVEF, alanine aminotransferase, serum creatinine, or creatinine clearance) were imputed using the population median. Prior to covariate testing, collinearity among candidate covariates was carefully evaluated using correlation analyses. The impact of each candidate covariate was first explored graphically by assessing their relationships with EBEs derived from base model (no covariates) for dimensionality reductions. Subsequently, the following covariates were screened for further analysis: age, body surface area (BSA), body mass index (BMI), cumulative DOX dose, concomitant enalapril use, and creatinine clearance. The relationship between precycle (baseline) GDF15 and prior DOX treatment cycles was explored graphically, showing increased baseline GDF15 with increased cumulative DOX doses (Figure 2), which supported evaluation of cumulative DOX dose as a covariate. Covariate selections were initially explored using Conditional Sampling for Stepwise Approach based on Correlation tests (COSSAC) approach in Monolix model building [19]. A covariate was considered to improve model fit if its inclusion led to a statistically significant decrease in the likelihood ratio test (e.g., ΔOFV > 3.84, assuming a chi‐square distribution with type‐I error of 0.05). COSSAC was used as an exploratory tool to inform covariate selection. Final covariate inclusion was guided by a full fixed‐effects modeling approach, prioritizing parameter plausibility, clinical interpretability, and physiological relevance. TABLE 1Patient characteristics.Continuous variablesMedian (range)Age (years)45.0 (28–58)Weight (kg)65.3 (54.5–140.4)Height (m)1.63 (1.52–1.88)BMI (kg/m2)25.3 (21.5–53.0)BSA (m2)1.71 (1.60–2.50)AST (IU/L)26.0 (18–33)ALT (IU/L)27.0 (8–62)Baseline LVEF (%)61.0 (48–69)Creatinine clearance (mL/min)120.7 (69.1–167.7)Categorical variables(Number of patients)Chemotherapy cycleCycle 1 (1)Cycle 2 (4)Cycle 3 (16)Cycle 4 (13)Abbreviations: ALT, alanine aminotransferase; AST, aspartate transaminase; BMI, body mass index; BSA, body surface area; LVEF, left ventricular ejection fraction. Patient characteristics. Abbreviations: ALT, alanine aminotransferase; AST, aspartate transaminase; BMI, body mass index; BSA, body surface area; LVEF, left ventricular ejection fraction. Relationship between baseline (precycle) GDF15 concentrations and chemotherapy cycle. Baseline (precycle) GDF15 concentrations stratified by chemotherapy cycle (Cycles 1–4), corresponding to increasing cumulative doxorubicin exposure (60, 120, 180, and 240 mg/m2, respectively). Each dot represents an individual patient observation, and bars represent median with interquartile range. Graph created using GraphPad Prism (Version 10.2.3). Model Evaluations Model stability and reliability were assessed based on successful model convergence (evaluated through trace plots of the OFV and parameter estimates), successful executions of the covariance step, and the physiological plausibility and imprecision of parameter estimates. Model fittings were evaluated using standard goodness‐of‐fit plots, including observed versus predicted plots (population and individual) and normalized prediction distribution error (NPDE) plots against predictions or time, and through prediction‐corrected visual predictive checks (pcVPCs) with 500 simulation replicates [20, 21]. Following successful model convergence and covariance steps, model comparison and selection were guided by Akaike and Bayesian information criteria (AIC/BIC), with smaller values indicating better fit. In addition, model convergence was further assessed using Monolix convergence assessment toolbox with five runs starting with different sets of initial estimates to ensure the stability of parameter estimates and robustness to initial conditions. Simulations Simulations were performed to compare the model‐simulated GDF15 and NT‐proBNP exposure profiles and to assess the prognostic performance of GDF15 as a biomarker for DOX‐induced cardiotoxicity. The final PK/PD models, including DOX and DOXOL, NT‐proBNP, and GDF15 concentrations were translated into mrgsolve in R [22]. The correctness of model translation was verified by comparing population predictions (PRED) and individual predictions (IPRED) generated by Monolix versus mrgsolve [23, 24]. Individual GDF15 and NT‐proBNP concentration–time profiles following DOX administration were simulated for 17 study participants using EBEs derived from the final PK/PD model. This is a deterministic simulation with all random effects set to zero [25]. This simulation was performed assuming a DOX dose of 60 mg/m2 per cycle. Each individual profile was simulated over 120 h using a dense time grid of every 0.01 h. Subsequently, exposure metrics, such as areas under the curve (AUCs), and the percentage change from baseline to the individual‐specific peak concentration for both biomarkers, were calculated using the simulated individual profiles. Time to peak concentration (T max) for GDF15 and NT‐proBNP was also derived from the simulated profiles for each individual and cycle. Paired comparisons of T max between biomarkers were performed using a nonparametric Wilcoxon signed‐rank test. Correlation analyses were then conducted to assess the relationships between GDF15, an emerging biomarker, and NT‐proBNP, the clinically established biomarker for cardiotoxicity monitoring, based on exposure metrics, in line with the second aim of the current study. Outliers were detected by iterative Grubbs' test [26] applied to baseline GDF15 concentrations. Once an outlier was identified, it was consistently excluded from all subsequent correlation analyses involving other parameters. A stochastic simulation considering IIV, but not RUV, was also conducted for the assessment of probability of target attainment (PTA). PTA is a metric that quantifies the likelihood of an individual's biomarker concentration exceeding a predefined threshold associated with a clinical outcome [27]. PTA was evaluated using literature‐derived cardiovascular risk criteria of GDF15: low risk (< 1200 pg/mL), moderate risk (1200–1800 pg/mL), and high risk (> 1800 pg/mL) [11, 28]. Briefly, we first generated a virtual cohort of 1000 patients by simulating body weight and height based on the distribution observed in the original dataset. An empirical mean vector and a covariance matrix of body weight and height were derived from the observed data set, and 1000 weight–height pairs were sampled from a truncated multivariate normal distribution using the tmvtnorm R package (version 1.7) [29]. For each virtual patient, BMI (kg/m2) was calculated and BSA (m2) was derived using the DuBois formula [30]. Simulated GDF15 concentrations were generated for these virtual patients at Day 14 postinfusion, representing the time point immediately preceding a subsequent chemotherapy cycle. PTAs were then evaluated according to these predefined GDF15 risk thresholds. Modeling Software PK/PD analyses were conducted using nonlinear mixed‐effects modeling in Monolix (version 2024R1; Lixoft, Antony, France). Model parameters were estimated using the stochastic approximation expectation–maximization (SAEM) algorithm [31], with an exploratory phase of 500–1500 iterations using auto stop criteria, followed by a smoothing phase of 200 iterations. Simulation was performed using mrgsolve (version 1.4.1). Data wrangling and visualization were performed in R (version 4.3.0, R Foundation for Statistical Computing, Vienna, Austria). Results Patients' Characteristics and GDF15 Concentration–Time Profiles Patient characteristics were previously reported and are summarized here in Table 1. The median age of patients was 45 years, with a median BSA of 1.71 m2. Patients received DOX at a dose of 60 mg/m2 per cycle, corresponding to cumulative doses of 60, 120, 180, or 240 mg/m2 after 1–4 cycles, respectively. Samples in this analysis were collected across different chemotherapy cycles (cycles 1–4), reflecting staggered patient enrollment into the randomized crossover study. Each patient contributed data from two consecutive crossover cycles, which occurred at different stages of their planned four‐cycle chemotherapy regimen. As a result, samples from 1, 4, 16, and 13 patients were obtained after cycles 1, 2, 3, and 4, respectively. A total of 271 GDF15 concentrations from 17 patients in two treatment cycles were used in PD model development. GDF15 concentrations started increasing as early as 2–3 h following DOX infusion. Figure 3 shows GDF15 concentration profiles during the early phase (0–5 h, Figure 3A) and across the full observation period (Figure 3B). Notably, patients who received higher cumulative doses of DOX (180–240 mg/m2; third and fourth chemotherapy cycles) exhibited higher mean precycle GDF15 concentrations, exceeding 1000 pg/mL, compared with those receiving lower cumulative doses (60–120 mg/m2; first and second chemotherapy cycles), whose GDF15 levels remained below 1000 pg/mL. Similarly, exploratory assessment of baseline (precycle) GDF15 concentrations across cumulative DOX dose levels demonstrated increased baseline (precycle) GDF15 concentrations with increasing cumulative DOX dose (Figure 2). Patients in the higher cumulative dose group also exhibited higher peak GDF15 concentrations compared to those in the lower dose group. Growth differentiation factor 15 (GDF15) concentrations following doxorubicin infusion. Individual patient responses (blue lines) and group mean (black lines) are shown for two cumulative dose groups (60–120 mg/m2 and 180–240 mg/m2). (A) Early response phase (0–5 h) and (B) full observation period (0–48 h). The PK/PD Model of DOX‐Induced GDF15 An indirect response model with four transit compartments (Figure 1C) adequately described the dynamic relationship between DOX exposure and GDF15 response. The model assumes that DOX stimulates GDF15 production (Figure 1C). The potential effect of DOXOL was also tested by adding its predicted plasma concentration in addition to C DOX, such that C drug = C DOX + C DOXOL in Equation (1); however, the inclusion of DOXOL did not improve model performance. Models with fewer or greater numbers of transit compartments were explored; however, the model with four transit compartments provided the best fit while precisely estimating model parameters, as indicated by lower AIC/BIC and relative standard errors of model parameters (Table 2). TABLE 2Parameter estimate for the pharmacodynamic model of GDF15.Parameter (unit)Fixed effect parametersRandom effect parametersShrinkage (%)Estimate (RSE)CV% (RSE) K tr (h−1)0.851 (0.792%) R 0 (pg/mL)744.866 (15.7%) IIV: 31.72 (18.7%) IOV: 11.04 (21.7%) 0.63 K out (h−1)1.012 (24.6%)48.86 (26.0%)9.858 E max 16.837 (11.5%)EC50 (μg/mL)0.0385 (21.3%)59.79 (20.0%)4.05 β 1 0.352 (32.8%) β 2 0.956 (38.2%)RUV: proportional0.156 (5.14%)corr(η_kout, η_EC50)−0.821 (19.7%) Note: Fixed effects include baseline GDF15 (R 0), transit rate constant (K tr), degradation rate (K out), maximal drug effect (E max), and half‐maximal effective concentration (EC50), β 1: covariate effect of cumulative DOX dose on R 0, β 2: covariate effect of BMI on R 0. Random‐effect variability (IIV and IOV) is reported as coefficient of variation (CV%), calculated using √(exp(ω 2) − 1). RUV follows a proportional error model and is calculated using √(σ 2). Shrinkage values for R 0 are reported for both IIV and IOV. corr(η_kout, η_EC50) denotes the correlation between the individual random effects on K out and EC50.Abbreviations: BMI, body mass index; CV%, coefficient of variation; IIV, inter‐individual variability; IOV, inter‐occasion variability; RSE, relative standard error; RUV, random unexplained variability. Parameter estimate for the pharmacodynamic model of GDF15. IIV: 31.72 (18.7%) IOV: 11.04 (21.7%) Note: Fixed effects include baseline GDF15 (R 0), transit rate constant (K tr), degradation rate (K out), maximal drug effect (E max), and half‐maximal effective concentration (EC50), β 1: covariate effect of cumulative DOX dose on R 0, β 2: covariate effect of BMI on R 0. Random‐effect variability (IIV and IOV) is reported as coefficient of variation (CV%), calculated using √(exp(ω 2) − 1). RUV follows a proportional error model and is calculated using √(σ 2). Shrinkage values for R 0 are reported for both IIV and IOV. corr(η_kout, η_EC50) denotes the correlation between the individual random effects on K out and EC50. Abbreviations: BMI, body mass index; CV%, coefficient of variation; IIV, inter‐individual variability; IOV, inter‐occasion variability; RSE, relative standard error; RUV, random unexplained variability. The stimulatory effect was best described by an Emax model. The parameter estimates of the final GDF15 PD model are summarized in Table 2. The estimated baseline concentration of GDF15 (R 0) was 744.9 pg/mL (95% confidence interval (CI): 550.7–1007.6) with IIV of 31.7% (95% CI: 21.9%–46.4%) and IOV of 11.0% (95% CI: 7.3%–16.7%). The mean transit rate constant (K tr) was estimated at 0.851 h−1 (95% CI: 0.838–0.864). The estimated degradation rate constant (K out) was 1.012 h−1 (95% CI: 0.637–1.606), while the maximum stimulatory effect (E max) was 16.84 (95% CI: 13.48–21.03), and the half‐maximal effective concentration (EC50) was 0.0385 μg/mL (95% CI: 0.0257–0.0578), respectively. Two significant covariate relationships were identified on R 0: cumulative DOX dose (β 1 = 0.352; 95% CI: 0.125–0.579) and BMI (β 2 = 0.956; 95% CI: 0.240–1.671), according to Equation (4): (4) R0,i=R0,pop.CumulativeDOXDosei60mg/m2β1.BMI25Kg/m2β2.eηIIV+ηIOV where R 0,i represents the individual baseline GDF15 concentration, and R 0,pop is the typical population baseline GDF15 value. Cumulative DOX Doseᵢ is the individual patient's cumulative DOX dose (mg/m2), with 60 mg/m2 as the reference cumulative dose used for normalization. BMI is the individual's BMI (kg/m2), with 25 kg/m2 as the reference BMI used for normalization. β 1 and β 2 are the covariate effect exponents describing the effect of cumulative DOX dose and BMI on baseline GDF15, respectively. ηIIV and ηIOV denote the IIV and IOV on R 0. A strong negative correlation (−0.821) was observed between the random effects of K out and EC50. The RUV was best described by a proportional error model. η‐Shrinkage for random effects was low (< 10%), indicating that the model provided reliable individual parameter estimates (i.e., EBE). Model Evaluations Standard goodness‐of‐fit plots suggest little evidence to reject the final model (Figure S1), and final model parameters were estimated with adequate precisions (%RSE < 50% for all parameter estimates, Table 2). pcVPC either in all patients (Figure 4A) or stratified by cumulative DOX dose categories 60–120 mg/m2 and 180–240 mg/m2, corresponding to chemotherapy cycles 1–2 and cycles 3–4, respectively, demonstrated that the final model adequately captured the observed GDF15 concentration–time profiles (Figure 4B). In addition, convergence assessment using five runs with different random initial estimates showed a relatively stable model convergence with respect to both objective function value and each model parameter (Figure S2). Prediction‐corrected visual predictive checks (pcVPCs) for GDF15 following doxorubicin infusion. (A) Overall pcVPC for all observations. (B) pcVPCs stratified by cumulative dose (left: 60–120 mg/m2; right: 180–240 mg/m2). Points are observed concentrations; blue solid lines are observed 5th/50th/95th percentiles; shaded bands represent the 90% prediction intervals from the model simulations: pink for the median, and blue for the lower (5th) and upper (95th) percentiles. Simulations Validation of model translation from Monolix to mrgsolve was performed using both population and individual predictions (PRED and IPRED). PRED and IPRED generated by mrgsolve closely aligned with those from Monolix, ensuring the accurate translation of model structure and parameters from Monolix to mrgsolve (Figure S3). Following the model translation, simulation was performed using EBE. Simulated concentration–time profiles using EBE aligned well with the observed ones, which confirmed the validity of EBE‐simulation (Data S2). Additionally, EBE‐simulation revealed distinct concentration–time profiles for GDF15 and NT‐proBNP following DOX administration across the two chemotherapy cycles (Figure S4). The median (range) time to T max for GDF15 was 8.8 h (7.9–9.5 h), whereas NT‐proBNP exhibited a delayed response with a median (range) T max of 34.8 h (13.2–49.8 h) postinfusion. The difference in T max between GDF15 and NT‐proBNP was statistically significant (p < 0.0001). To further compare the model‐simulated GDF15 responses with NT‐proBNP in the same patient populations, several GDF15 and NT‐proBNP exposure metrics, including peak, AUC, and %change from baseline (precycle) to T max were calculated, and the relationships between them were investigated (Figure 5). Outliers from one patient in two cycles, who had extraordinarily high baseline values of GDF15, were excluded based on iterative Grubbs' test for outlier detection. At baseline, concentrations of GDF15 and NT‐proBNP were strongly and significantly correlated (r = 0.631, p < 0.05) (Figure 5A). Following DOX administration, NT‐proBNP and GDF15 peak concentrations (r = 0.086, p = 0.639, Figure 5B) and AUC (r = 0.110, p = 0.549, Figure 5C) exhibited weak but consistently positive correlation (Figure 5B). However, the percent change from baseline in GDF15 versus NT‐proBNP (Figure 5D) showed moderately positive significant correlations (r = 0.362, p = 0.0414). Similarly, the percent change in AUC for GDF15 versus NT‐proBNP demonstrated positive, though not statistically significant, correlation (r = 0.286, p = 0.112) (Figure 5E). Relationships between simulated GDF15 and NT‐proBNP responses. (A) Baseline NT‐proBNP vs. baseline GDF15. (B) Peak NT‐proBNP vs. peak GDF15. (C) NT‐proBNP AUC vs. GDF15 AUC. (D) percent change from baseline in NT‐proBNP vs. GDF15. (E) Percent AUC change in NT‐proBNP vs. percent AUC change in GDF15; outliers are highlighted in red and were excluded from the regression fit. Each point represents one individual; blue lines represent linear regression using the ordinary least square method. A stochastic simulation with 1000 virtual patients was also performed for the assessment of PTA using predefined cardiotoxicity criteria of GDF15 [11, 28]. The generated virtual population demonstrated height, weight, BMI, and BSA distributions similar to those of the original study population (Table S1). In this simulation, Day‐14 GDF15 concentrations, representing the final time before the subsequent chemotherapy cycle, demonstrated a dose‐dependent increase with increasing cumulative DOX dose (60, 120, 180, 240 mg/m2) (Figure 6A). Across the simulated population, 64.7%, 26.6%, and 8.7% of patients were classified as low, moderate, and high risk of DOX‐induced cardiotoxicity, respectively (Figure 6B). Most low‐risk patients (< 1200 pg/mL) had received lower cumulative doses, with 34.6% receiving 60 mg/m2. In contrast, a large proportion of patients receiving 240 mg/m2 were predicted to fall into the high‐risk category, representing 48.3% of the total population of this category. Relationship between accumulated doxorubicin dose and GDF15 concentrations. (A) Simulated GDF15 concentrations from 1000 virtual patients at Day 14 post‐infusion, representing the last day before the subsequent chemotherapy cycle, are plotted against accumulated doxorubicin dose (60, 120, 180, and 240 mg/m2). Each point represents an individual patient, with a fitted linear regression line shown in red. Horizontal dashed lines indicate predefined GDF15 thresholds for cardiovascular risk categories [11, 28]: low risk (< 1200 pg/mL), moderate risk (1200–1800 pg/mL), and high risk (> 1800 pg/mL). (B) Percentage distribution of simulated patients across GDF15 risk categories, stratified by accumulated dose, derived from the simulation in (A). The stacked bars show both the proportion of patients within each risk category and the contribution of different accumulated dose groups to the overall distribution within each risk category. Discussion Anthracyclines are effective chemotherapeutic agents widely used in the treatment of multiple malignancies, including breast cancer, lymphomas, and various solid tumors [1]. However, their clinical utility is limited by dose‐dependent cardiotoxicity [2]. Circulating biomarkers, in addition to echocardiography, have been shown to enhance early detection of cardiotoxicity and improve prediction of late‐onset cardiac dysfunction [32, 33]. Among emerging biomarkers, GDF15 has gained attention as a promising indicator of subclinical myocardial stress and a potential early marker of anthracycline‐induced cardiotoxicity [4]. GDF15 is a stress‐responsive cytokine and has emerged as an important biomarker of cardiovascular stress, inflammation, and injury [7]. A growing body of evidence highlights GDF15 as a valuable biomarker for cardiovascular risk across diverse disease states. In patients with atrial fibrillation receiving oral anticoagulation, GDF15 predicts major bleeding, adverse cardiovascular events, and all‐cause mortality [8]. In ischemic and nonischemic cardiomyopathy, elevated GDF15 levels are associated with an increased risk of all‐cause mortality and heart failure hospitalization [10]. Moreover, in acute myocardial infarction, higher GDF15 concentrations are linked to a greater likelihood of sudden cardiac death within the first 24 h [9]. A recent meta‐analysis further demonstrated that GDF15 consistently provides prognostic prediction regarding the risk of cardiovascular death and heart failure hospitalization in different types of atherosclerotic cardiovascular disease [11]. Importantly, GDF15 has also been proposed as a promising biomarker for chemotherapy‐induced cardiotoxicity, independently predicting cardiovascular risk [4, 12]. However, there is currently no systemic model characterizing the temporal dynamics of DOX exposure and GDF15 response, limiting our understanding of the prognostic utility of GDF15 and how it compares with existing circulating cardiovascular biomarkers, such as NT‐proBNP. Therefore, in the present study, we employed a model‐based PK/PD approach to characterize the temporal exposure‐response relationship between DOX and GDF15, and performed simulations to further evaluate the prognostic value of GDF15. To our knowledge, this is the first study to characterize DOX‐induced GDF15 dynamics via developing a PK/PD model. In the present study, a PK/PD model characterizing the relationship between DOX exposure and GDF15 response was developed via expanding a previously published model [13]. An indirect response model with stimulatory effect using Emax model on GDF15 production adequately captured the observed GDF15 profiles while preserving model parsimony. The model incorporates four transit compartments to capture the observed biological delay in GDF15 increase following DOX administration. Relatively high IIV (31.72%–59.79%) was observed for the PD parameters R 0, K out, and EC50. Baseline GDF15 level (R 0) was significantly and positively associated with two key covariates: BMI and cumulative DOX dose. This is consistent with previous reports showing that circulating GDF15 is higher in individuals with obesity and correlates positively with BMI [34]. Similarly, plasma GDF15 was significantly higher in obese patients compared to nonobese patients undergoing heart surgery [35]. Mechanistic studies demonstrated that GDF15 is released in response to metabolic and inflammatory stress signal in obesity [36], serving as a compensatory signal to limit further weight gain and metabolic dysfunction, by stimulating satiety and promoting weight loss through activation of the GDF15–GFRAL pathway in the hindbrain [37, 38]. In addition, we observed the covariate effect of cumulative DOX dose on baseline GDF15 levels (R 0) in our model. This aligns with clinical evidence indicating that anthracycline exposure increases circulating GDF15. In a prospective cohort study, median plasma GDF15 increased from approximately 690 pg/mL at baseline to 1259 pg/mL at the end of DOX‐based chemotherapy, reflecting a cellular stress response to cumulative anthracycline exposure [4]. Notably, GDF15 levels remained elevated 3–6 months after treatment, despite a partial posttherapy decline [4]. Similarly, in children and adolescents undergoing anthracycline therapy, serum GDF15 levels posttreatment were significantly higher than pretreatment baseline values [39]. Higher DOX exposure is associated with a greater risk of cardiac injury, and elevated GDF15 levels were associated with this risk. In a longitudinal study of breast cancer patients (treated with DOX followed by taxanes ± trastuzumab), increases in GDF15 persisted up to 15 months and were significantly associated with subsequent cardiotoxicity events [40]. Cumulative DOX dose was incorporated in the model as a continuous covariate on baseline GDF15 (R 0) to account for treatment history–associated differences. Clinically, DOX is administered intravenously in cycles, typically every 2–3 weeks, including dose‐dense regimens of 14 days as in the current study. Although DOX has a relatively short half‐life (~20–48 h), with negligible systemic drug concentrations expected at the start of each subsequent cycle, prior cycles appear to exert a persistent or partially irreversible effect on baseline GDF15 concentrations, as supported by our exploratory analysis (Figure 2). We observed a clear trend in which higher cumulative doses were associated with elevated baseline GDF15 levels. It is crucial to note that cumulative dose was applied exclusively to baseline GDF15 (R 0) and was not used to drive the acute exposure‐response relationship, which was governed solely by individual DOX concentration–time profiles from the current cycle. Similar to DOX/NT‐proBNP PK/PD model, concurrent enalapril administration did not show a significant effect on the GDF15 PD parameters, which is consistent with a previous study reporting no effect of enalapril on changes in GDF15 levels over time [41]. Importantly, results from the parent drug–drug interaction study demonstrated that enalapril co‐administration did not alter DOX systemic exposure, with no clinically or statistically meaningful differences observed in AUC, C max, or half‐life. In addition, DOX exposure was consistent across treatment cycles, as the same dose (60 mg/m2) was administered in each cycle. Together, these findings indicate that enalapril did not introduce PK confounding and support the interpretation that observed GDF15 dynamics reflect DOX‐related effects rather than enalapril‐ or cycle‐dependent changes in exposure. Although circulating GDF15 levels are known to increase with age and have been associated with biological aging and age‐related pathophysiology [42], we did not identify a statistically significant covariate effect of age in our cohort due to the relatively limited age range of participants (28–58 years). Using a population simulation (stochastic PTA simulation), we simulated Day‐14 GDF15 concentrations, a clinically relevant timepoint prior to subsequent chemotherapy, in 1000 virtual breast cancer patients across four cumulative anthracycline dose levels (60, 120, 180, and 240 mg/m2). We applied GDF15‐based cardiovascular risk categories according to previously published thresholds [< 1200 pg/mL (low risk), 1200–1800 pg/mL (moderate risk), and ≥ 1800 pg/mL (high risk)] [11, 28]. The simulations revealed a cumulative DOX dose‐dependent increase in GDF15 levels. We found that 8.7% of the simulated cohort fell into the high‐risk group, while 26.6% and 64.7% were classified as moderate and low risk, respectively. Notably, patients receiving 240 mg/m2 of cumulative DOX comprised 48.3% of the high‐risk group. The 8.7% high‐risk estimate closely mirrors the reported incidence of overt anthracycline‐induced cardiotoxicity, which ranges from 2% to 9% [3], supporting the translational value of GDF15‐based risk modeling. Importantly, the demographics of the virtual cohort were comparable to reported breast cancer cohorts treated with anthracyclines [43, 44], including age and BMI distributions observed in larger real‐world studies (Table S2). This supports the clinical relevance of the simulated population for exploratory risk stratification analyses. Our population simulation results, showing a positive relationship between baseline GDF15 levels and cumulative DOX doses (Figure 6A), are supported by our exploratory analysis and clinical evidence. In our study, baseline (precycle) GDF15 concentrations increased with increased DOX treatment cycles (cumulative DOX dose), as illustrated in Figure 2. In a separate study, baseline GDF15 levels significantly differed between patients receiving higher versus lower cumulative anthracycline doses [41]. Baseline GDF15 and changes in GDF15 at both 1 month and 8 months were associated with subsequent mortality and CV events. Notably, each 20% increment in baseline GDF15 was linked to a higher risk of mortality, the composite endpoint of cardiovascular death or heart failure hospitalization, and heart failure‐related death [41]. Together, these clinical and simulation findings highlight the value of GDF15 as a predictor of cardiotoxicity in patients treated with anthracyclines and support its potential use in dose‐dependent risk assessment. In our EBE‐based simulations, GDF15 levels increased rapidly following DOX infusion, peaking approximately 8 h postinfusion, while NT‐proBNP exhibited a delayed response, reaching peak levels around 34 h. Despite both biomarkers showing elevated responses to treatment, the correlation between their peak values and AUCs was weak and statistically nonsignificant. This weak correlation may be due to the small sample size and the distinct biological pathways governing the release of each biomarker. While GDF15 is a stress‐responsive cytokine that rises in response to early oxidative and inflammatory cellular stress [7], NT‐proBNP is primarily released due to myocardial stretch and hemodynamic overload [6], which may occur later in the pathophysiological process. Notably, the percent changes from baseline of GDF15 and NT‐proBNP showed a significant moderate positive correlation, and a similar positive correlation was observed for the percent AUC change for both biomarkers. These correlation analyses are exploratory in nature and should be interpreted as hypothesis‐generating, given the limited sample size of the current study and require confirmation in larger, prospectively designed cohorts. Consistent with our findings, prior studies have demonstrated that GDF15 and NT‐proBNP are only moderately correlated (r~0.3–0.5) but provide complementary prognostic value. In acute heart failure, each biomarker independently predicted adverse cardiovascular outcomes, and their combination improved mortality prediction compared to either alone [45, 46]. Similarly, in studies of vascular brain injury and dementia, both markers were moderately correlated and independently linked to vascular brain injury and cognitive decline [47]. Importantly, in patients receiving anti‐HER2 therapy ± anthracyclines, GDF15 demonstrated a moderate correlation with NT‐proBNP and independently predicted cardiotoxicity risk, with an area under the receiver operating characteristic (ROC) of 0.803 [12]. This study has several limitations. First, the analysis was based on a small sample size (17 patients) and represents a secondary, exploratory analysis of a clinical trial that was not originally designed for biomarker modeling. Consequently, the study was not powered for definitive biomarker evaluation, and the findings should be interpreted as hypothesis‐generating, requiring confirmation in larger, well‐powered, prospectively designed cohorts. In addition, PK/PD sampling was conducted during only two chemotherapy cycles per patient as part of the randomized crossover design. Although these two consecutive cycles occurred at different stages of the planned four‐cycle chemotherapy regimen across the study population, allowing variability in prior cumulative DOX treatment history, the study design does not provide full longitudinal sampling across all treatment cycles within the same individuals. Therefore, the current analysis primarily characterizes the acute PD response of GDF15 to DOX rather than the complete trajectory of cumulative or potentially irreversible biomarker changes across the entire chemotherapy course. In addition, the virtual patient cohort used for simulations was generated based on the demographic characteristics of this small clinical study. Although the resulting virtual population closely resembled published breast cancer cohorts treated with DOX (Table S2), it cannot fully capture the inherent heterogeneity of real‐world breast cancer populations. Furthermore, cumulative DOX dose in this dataset was limited to four observed levels (60–240 mg/m2), reflecting standard clinical dosing. Although modeled as a continuous covariate, this restricted exposure range may limit extrapolation beyond the studied doses and warrants validation in larger cohorts with broader exposure variability. In addition, no clinical cardiotoxicity endpoints (e.g., LVEF decline) were collected, which limits the ability to determine whether modeled changes in GDF15 or NT‐proBNP are predictive of actual clinical outcomes. Slight overprediction at higher concentrations was observed in the NPDE versus population predictions, likely due to limited data density in this range and therefore could not be fully resolved within the current dataset. Nonetheless, overall goodness‐of‐fit diagnostics and predictive performance were acceptable for the primary objectives of this exploratory analysis. In conclusion, we developed the first PK/PD model characterizing DOX‐induced GDF15 production. Our analysis identified cumulative DOX dose and BMI as significant covariates influencing baseline GDF15 levels. Simulations using the developed model demonstrated a cumulative DOX dose‐dependent increase in GDF15 levels and enabled cardiovascular risk stratification based on predefined GDF15 thresholds. Comparisons with NT‐proBNP demonstrated that GDF15 responds earlier to cellular stress, highlighting its potential as a complementary biomarker for anthracycline‐induced cardiotoxicity. However, given the exploratory nature of this secondary analysis and the small study cohort (n = 17), these findings should be interpreted as hypothesis‐generating. Larger, prospectively designed studies are required to validate the temporal dynamics of GDF15 and to confirm its clinical utility for early detection and risk stratification of anthracycline‐induced cardiotoxicity. Author Contributions M.S.D., P.A.J., M.N.K., B.N.Z., and S.C. wrote the manuscript. A.H.B., P.A.J., M.N.K., B.N.Z., and S.C. designed the research. M.S.D., A.H.B., P.A.J., M.N.K., B.N.Z., and S.C. performed the research. M.S.D., and S.C. analyzed the data. Funding The authors have nothing to report. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Data S1: cts70595‐sup‐0001‐DataS1.docx. Figure S1: Goodness‐of‐fit plots and residual diagnostic plots for the pharmacodynamic model of DOX‐induced growth differentiation factor 15 (GDF15) production. (A) Observed GDF15 concentrations versus population predictions (PRED). (B) observed GDF15 concentrations versus individual predictions (IPRED). The solid line represents the line of identity. (C) Normalized Prediction Distribution Errors (NPDE) versus time. (D) NPDE versus population predictions. Figure S2: Convergence assessment. Trace plots of parameters estimates and the objective function value with respect to the iterations of stochastic approximation expectation–maximization. Color represents runs with different random initial estimates. Figure S3: Comparison of population and individual predictions (PRED and IPRED) of doxorubicin (DOX), doxorubicinol (doxol), NT‐proBNP, and GDF15 generated by Monolix vs. translated mrgsolve models. Panels A–D show PRED and IPRED of DOX and doxol; panels E–F show PRED and IPRED of NT‐proBNP; and panels G–H show PRED and IPRED of GDF15. Figure S4: Simulated concentration–time profiles using EBE of GDF15 and NT‐proBNP concentrations. Simulations were performed using EBE for the study patients across the two chemotherapy cycles. Each colored line represents the trajectory of an individual patient. (A) Simulated GDF15 concentrations over time following doxorubicin infusion for Cycle 1 (left) and Cycle 2 (right). (B) Simulated NT‐proBNP concentrations over time following doxorubicin infusion for Cycle 1 (left) and Cycle 2 (right). Table S1: Baseline characteristics of the original and virtual patient populations. Table S2: Comparison of demographic characteristics between the current study, virtual patient cohort, and selected breast cancer cohorts. Data S2: Comparison of observed and empirical Bayes estimate (EBE)‐simulated concentration–time profiles.